autogait: a mobile platform that accurately estimates the distance walked 1 dae-ki cho min mun,...
TRANSCRIPT
AutoGait: A Mobile Platform that Accurately Estimates the Distance Walked
1
Dae-Ki Cho
Min Mun, Uichin Lee, William J. Kaiser, Mario Gerla
Outline
Motivation Intro to AutoGait & Our Approach Prototype Implementation Experiment Results Summary and Future Work
2
Pedometers & Applications
Can be applied to various applications Pedestrian Dead Reckoning for Indoor Navigation and Outdoor
Trajectory Tracking RF-based localization requires infrastructure 3G localization is not accurate GPS doesn’t work indoor environment and consumes lots of battery
Activity/Health Monitoring Monitoring Ambulatory Activity
etc.
3
Limitations of existing approaches
4
Manual calibration : Inconvenient, Erroneous, Tedious
Use of constant stride length : Seriously biasing estimation results
Why accuracy is important?
Small error in each step could result in a huge difference in estimating total distance walked Experiment results: actual distance walked:
400m, vs. Omron pedometer: 496.3m in slow speeds and 341.6m in fast speeds
[5] K. De Cocker, G. Cardon, and I. De Bourdeaudhuij, “Validity of the inexpensive Stepping Meter in counting steps in free living conditions: a pilot study”, Br J Sports Med, vol. 40, no. 8, pp. 714–716, 2006.
For some applications like indoor navigation/pedestrian dead reckoning system, a few meters of error could account for location misprediction.
5
The Goal of AutoGaitA mobile platform that autonomously discovers a user’s walking profile when
the user walks outdoors by utilizing the mobile’s GPS and the step detector
accurately estimates the distance walked using the calibrated walking profile without GPS
6
Walking Profile – Variable Stride Length
7
Step Frequency
Strid
e Le
ngth
Physiologists found there is a linear relationship between step frequency and stride length
Estimating Distance Walked
1. Measure the step frequency (fi) for each step2. Calculate the stride: si = α×fi+β3. Add si to the cumulative distance walked (D): Dnew = Dold + si
8
Step Frequency
Strid
e Le
ngth
f
s Walking profile
s = α×f+β
Problem when calibrating with GPS in Mobile Device
9
(Latitude1, Longitude1) (Latitude2, Longitude2)
Stride lengthStep freq
GPS in mobile devices have the error range of 5 to 10 meter.
Step Frequency
Strid
e Le
ngth
Stride lengthStep freq
Our Approach
10
For each cluster, we compute average stride length (SL) and step frequency (SF).
(SL 1, SF 1)
(SL 3, SF 3
)
(SL2, SF2)
Step Frequency
Strid
e Le
ngth
Straight Line Identifier
Segmentation
Smoothing
3-step GPS data filtering
Constant Stride length in the beginning
Step 1: Segmentation (Pre-process)
11
Step Frequency (Hz) = 1/Step Interval (sec)
Step interval becomes larger when a user stops.
✓ Immobility Detection
• Remove edge of two consecutive GPS readings if it contains a step interval is greater than (mean + 3 X standard deviation) of the total step interval
Step 1: Segmentation (Pre-process)
• Environmental obstacles generate sudden jumps in GPS traces
• Remove noisy samples1. Speed between two consecutive GPS readings > (mean +
2*standard deviation) of the total speed
2. Remove sub-segment if it contains a few GPS coordinates
12
Speed becomes larger when a GPS coordinate jumps due to noise.
Noisy SamplesClear Samples
✓ Unrealistic Movement Detection
Step 2: Smoothing
Convolution is used for the smoothingEach GPS coordinate in the segment is smoothed
with only its neighbors
13
(lat2, lng2)
…(lat1, lng1)(lat3, lng3)
(lat4, lng4)
(latn-3, lngn-3)
(latn-2, lngn-2)
(latn-1, lngn-1)
(latn, lngn)
(conv_lat1, conv_lng1) = ((lat1+lat2+lat3)/3, (lng1+lng2+lng3)/3)
(conv_latn-2, conv_lngn-2) = ((latn-2+latn-1+latn)/3, (lngn-2+lngn-1+lngn)/3)
…
Step 2: Smoothing
14
However - Some of noisy GPS readings still remain Smoothing makes sharp corners dull and round
Step 3: Straight-Line Identifier
Heading Change-based filtering: focus only on walking patterns in straight-line roads
Ci: Δ angle between edge (P1, P2) and edge (P1, Pi+2)
Find max i where C1…Ci-1 < MT, Ci< ET
15
P2
P3
Pi+2Pi+1
..….
End-point Threshold (ET)
Pi+4
Middle-point Threshold (MT)
P4
P1
Good straight-line segment
Prototype Implementation
• Nokia N810 using Linux Python and GPS• Pedometer Implementation
– Pressure Sensors in UCLA SmartShoe platform – MicroLEAP acquires sensor data and transfers it using
Bluetooth
16
Linear Relationship Verification
17
Measured on a treadmill Two hundred steps per each speed A line generated using the linear regression Sample points are very close to the line
Speed (mph) Frequency (Hz) Stride length (cm)
1.0 0.51 43.90
1.5 0.65 51.79
2.0 0.78 58.07
2.5 0.91 62.55
3.0 0.10 68.37
3.5 1.07 74.17
4.0 1.14 80.00
4.5 1.22 83.93
Step Frequency (Hz)
Str
ide L
ength
(cm
)
Identifying Straight-Line Segments
18
(a) Raw GPS Data (b) Effect of Segmentation and Smoothing
(c) Heading Change based Filtering
A dataset obtained by walking near the UCLA campus 6 routes (trials), 26 straight lines were detected
Linear Profile Calibration As number of samples increase, the slope variations gradually converge. We terminate the calibration process when the slope variation
sequentially stays within ±1o over multiple time periods. Once the calibration is done, the system turns off the GPS module and
uses the calibrated profile to estimate the distance walked.
19
Step Frequency (Hz)
Str
ide L
ength
(m
)
Number of Samples
Δ S
lope A
ngle
(o)
Effectiveness of GPS Filtering
20
AutoGaitTreadmillRaw GPS
Step Frequency (Hz)
Str
ide L
ength
(cm
) The lines generated by AutoGait and Treadmill are following similar slope AutoGait is above the Treadmill because the stride length increases when the user walks
on the ground compared to the treadmill H. Stolze, J. P. Kuhtz-Buschbeck, C. Mondwurf, A. Boczek-Funcke, K. Jhnk, G. Deuschl, and M. Illert, “Gait analysis during treadmill
and overground locomotion in children and adults,” Electroencephalography and Clinical Neurophysiology/Electromyography and Motor Control, vol. 105, no. 6, pp. 490 – 497, 1997.
The raw GPS significantly overestimates the stride length due to the noise
Validation: Field Test
AutoGait outperforms the constant stride length-based method both at slow speeds and at fast speeds.
Considering the state of arts such as Nike+Apple Shoe or Omron pedometer uses the constant stride length, AutoGait can enhance the accuracy.
21
Speed Slow Moderate Fast
Distance (m) 400 800 400
Lap Time 9:56 11:52 3:45
# of Steps (Ground truth) 718 1192 488
AutoGaitEst Dist (m) 395.9 795.4 396.3
Error Rate 1.02% 0.58% 0.93%
Constant Stride Length (70 cm)
Est Dist (m) 502.6 834.4 341.6
Error Rate -25.7% -4.3% 14.6%
Testing on Multiple Users
Participant A B Cα (SLL) 0.453 0.064 0.539β (SLL) 0.23 0.612 0.2156
Est Dist (m) 1577.5 1579.4 1616.9Error Rate -1.41% -1.29% 1.06%
Three people participated (A, B, and C) Walking Profile Calibration:
Casually walked around the UCLA campus α and β are different for individuals – the profile should
be personalized Validation:
Walked four laps on a track (i.e., 1.6 km)
22
Summary and Future work
Developed a mobile platform that autonomously find the variable stride length, which can be used to accurately estimate the distance walked
Implemented prototype using Nokia N810Extensive experiments significantly lower the error
rates, achieving more than 98% accuracy in our testbed scenarios
Future work Outdoors and indoors detection Consideration of physiological factors
o In case of runningo Walking uphill and downhill - Altitude
23
Appendix
25
AutoGait Architecture
Step Detector
GPS Training Dataset
Stride Length Lookup (SLL)
Trigger Re-calibration
Straight Line Identifier
GPS enabled Mobile
Record Activity
GPS data filtering and calibration module
Pedometer module
Segmentation
Smoothing
Input: Step Frequency Output: Updating SLL
26
Validation: Field Test (1)
27
High error rates for RG and RGO SM underestimates the distance traveled The sum up method performs two times better than the end-to-end method in the track
test The sum up method works four times better than TM in the track test, but the TM
performs three times better than the sum up method in the treadmill. Implying the SLL should be discovered outdoors for a better estimation
Terms
Δt: Step interval
Step Frequency = 1/Step interval
Stride length
Effectiveness of GPS Filtering
The lines generated by two HC methods are above the Treadmill but they are following similar slope
Two HCs are above the Treadmill because the stride length increases slightly when the user walks on the ground compared to when the treadmill
The raw GPS significantly overestimates the stride length
29
Benchmark StudiesOmron Pedometer (HJ-720ITC) Nokia Step Counter
Speed Slow Moderate Fast Slow Moderate Fast
Found Steps
709 1190 488 266 1051 456
Est Dist (m) 496.3 833 341.6 181.7 717.8 311.4
Error Rate -24.08% -4.13% 14.6% 54.6% 10.3% 22.1%
400 m Test Without Calibration With Calibration
Speed Slow Moderate Fast Slow Moderate Fast
Lap Time 8:41 5:08 3:26 8:21 4:59 3:34
Est Dist (m) 160 460 390 290 410 360
Error Rate -60% 15% -2.5% -27.5% 2.5% -10%
30
Accelerometer based pedometer may not detect the steps at slow speeds. The Nike shoes use a constant stride length for estimating the distance
walked.
Discussion & Future Studies
Low-acceleration problem of Accelerometer-based pedometers
Outdoors and Indoors detectionAutoGait on Indoor Navigation SystemsConsideration of Physiological Factors
31